Training Dense Retrievers with Multiple Positive Passages

Researchers have developed a new training method for retrieval systems that uses multiple positive examples instead of just one, overcoming a major bottleneck in AI accuracy. The study tested four approaches across major benchmarks like MS MARCO and BEIR, finding the LSEPair loss consistently performed best. This method leverages LLMs to generate richer training data, potentially solving the false-negative problem that has limited retrieval-augmented generation (RAG) systems. The findings provide clear design principles for building more reliable AI assistants.